Evaluation of AI-Enabled POCUS Versus Standard Echocardiography in Hospitalized Patients

Evan Baum, MD*; John Kugler, MD; Michelle Lo, MD; Caitlin Parmer-Chow, MD; Vishnu Nair, MD; Andre Kumar, MD MEd

Department of Medicine, Stanford University School of Medicine, Stanford, CA, USA

*Corresponding Author: Dr. Evan Baum (email: evanbaum@stanford.edu)


Download article PDF – POCUS Journal 2026;11(2):71-79

DOI: https://doi.org/10.24908/pocusj.v11i02.21227

Supplementary Material: S1


Abstract

Background: Artificial intelligence (AI) may improve ultrasound image acquisition and interpretation, but there have been few validation studies comparing AI-enabled machines to traditional measurement workflows. In this diagnostic comparison study, we evaluated AI-enabled point of care ultrasound (POCUS) compared to standard transthoracic echocardiography (TTE) measurements. Methods: Hospitalized adults at a single academic center who underwent both AI POCUS and standard TTE scans within 7 days of one another were eligible for inclusion. The AI POCUS software used convolutional neural network models to estimate left ventricular ejection fraction (LVEF), right ventricular fractional area change (RVFAC), valvular regurgitation, valvular stenosis, and inferior vena cava (IVC) diameter based on standard cardiac views. Agreement between AI POCUS and TTE was assessed using weighted Cohen’s κ and exploratory Bland-Altman analyses. Results: Seventy patients were included, with a median of 1 day (IQR 1–2 days, range 0–6 days) between AI POCUS and TTE examinations. AI POCUS demonstrated substantial agreement with TTE for LVEF (κ = 0.75, 95% CI 0.42–0.92). Agreement was lower for RVFAC (κ = -0.003), IVC diameter (κ = 0.382), aortic regurgitation (κ = 0.541), mitral regurgitation (κ = 0.438), tricuspid regurgitation (κ = 0.400), and aortic stenosis (κ = 0.086). Using a dichotomized threshold of LVEF <50% versus ≥50%, AI POCUS demonstrated 80.0% sensitivity and 88.6% specificity for detecting reduced LVEF. Conclusions: AI POCUS demonstrated substantial agreement with TTE for the estimation of LVEF. Agreement was lower for other echocardiographic parameters. These findings support the potential of AI POCUS for assessing left ventricular function.
 

Introduction

Approximately 56 million people are affected by heart failure worldwide [1,2]. Echocardiography, the gold-standard test for evaluating heart failure [3], is available in less than 7% of healthcare facilities in lower- and middle-income countries [4], contributing to delayed diagnosis and treatment for heart failure for millions worldwide [2].

Point of care ultrasound (POCUS) can help address shortages and barriers to accessing formal echocardiography [5]. POCUS differs from traditional echocardiography because the clinician obtains and interprets images in real time at the bedside, rather than a sonographer acquiring images that a cardiologist later interprets [5]. Previous randomized trials have demonstrated that POCUS can improve time to diagnosis among patients presenting with symptoms concerning for decompensated heart failure [6], while others have demonstrated improved quality of care among hospitalized patients with heart failure who received POCUS [7].

Recently, artificial intelligence (AI) algorithms have been integrated into POCUS systems, aiding in image acquisition and image quality [8–10]. AI can also aid in interpretation of both formal echocardiograms and POCUS examinations with a high degree of accuracy [11], including for measurements of left ventricular ejection fraction (LVEF) [12,13].

While there has been demonstrated utility of AI-enabled POCUS to assist with diagnosing heart failure [8–10,13], there is a critical need to further validate the accuracy of emerging AI systems in a clinical context, including their utility in categorizing disease severity compared to reference standards [14]. Notably, among all Food and Drug Administration (FDA) approvals for AI-enabled devices, fewer than 10% of the applications reported data from prospective clinical studies [15]. Validation studies are vital to improve provider and patient trust in AI and to facilitate larger systemic adoption to address current delivery limitations for heart failure [14].

The objective of this study was to assess the agreement of measurements between standard transthoracic echocardiography (TTE) and an AI POCUS software product that is FDA-approved and commercially available for clinical environments. We hypothesized that AI POCUS would show substantial agreement with TTE for assessments of systolic function and at least moderate agreement for assessments of valvular function.

Methods

Setting and Participants

This single-arm, prospective diagnostic comparison study included 70 adult inpatients at a single tertiary medical center in California, USA, between April and October 2025. Consecutive adult patients ≥18 years were identified from a convenience sample of individuals who were admitted to medical-surgical or step-down units and received a TTE for routine clinical care. Patients were eligible for the study if they received an AI POCUS examination within 7 days of their TTE. Patients were excluded if they had history of congenital heart disease, mechanical valve replacement, or did not receive an AI POCUS scan within 7 days of TTE. The Stanford University Institutional Review Board approved the study (IRB Protocol #77758) and waived the requirement for informed consent.

Scanning Protocol and AI Software

Comprehensive TTEs were performed by cardiac sonographers in accordance with American Society of Echocardiography (ASE) guidelines [16]. All TTEs were interpreted by a board-certified cardiologist credentialed at the medical center for echocardiography, as part of the patient’s standard care. POCUS exams were performed by three expert users (EB, JK, AK) using a Philips Lumify S4-1 Phased Array probe on a Samsung tablet. For the POCUS examination, four standard views were obtained: parasternal long (PLAX), parasternal short (PSAX), apical four-chamber (A4C), and subcostal inferior vena cava (IVC). Views were obtained in both B-mode and with colour Doppler and archived.

All POCUS images were then analyzed on an AI platform, AISAP, which is an FDA-approved software application for POCUS and AI. As part of its regulatory approval, AISAP was validated on a large patient database and demonstrated high diagnostic performance in retrospective studies for classifying LVEF, right ventricular fractional area change (RVFAC), and valvular disease [17,18].

The AISAP model was internally validated using over 140,000 TTE studies acquired between 2007 and 2021 [17]. All studies adhered to ASE standards and were labelled using final clinical interpretations by board-certified cardiologists. The model was trained via supervised learning to assess severity classifications assigned by clinicians and validated on POCUS cardiac images [18].

The diagnostic model computes LVEF, RVFAC, mitral regurgitation (MR), tricuspid regurgitation (TR), aortic regurgitation (AR), aortic stenosis (AS), and IVC diameter. It does so from the following views: PLAX with and without colour Doppler over the aortic valve, PSAX at the level of the aortic valve, A4C with and without colour Doppler over the tricuspid and mitral valves, and standard subcostal IVC views. The software provides real-time feedback on acceptability of image quality. During processing, if the model determines that the views obtained do not meet sufficient diagnostic quality for interpretation, it will not provide an estimate. The model does not rely on guideline-based measurements such as proximal isovelocity surface area (PISA), vena contracta width, or effective regurgitant orifice area (EROA) for valvular assessments. Instead, it infers severity based on learned visual patterns and assigns a probability to each severity class (none, mild, moderate, severe) and selects the most likely category [18]. A full description of the model is available in the Supplementary Materials.

Statistical Analysis

Results from AISAP and TTE reports were classified using predefined categories. LVEF was classified into four categories based on the American Heart Association (AHA) guidelines for heart failure: ≥50%, 41–49%, 26–40%, and ≤25% [3,19]. Notably, the ≤25% cutoff for LVEF is based on the 2024 AHA imaging guidelines [19]. RVFAC was categorized based on the 2025 ASE guidelines: >35%, 30–35%, 23–29%, and ≤22% [20]. Valvular abnormalities, including MR, TR, AR, and AS were assigned based on AHA classifications of valvular disease: none to trace, mild to moderate, moderate to severe, and severe [21]. IVC diameter was categorized as ≤2.1 cm or >2.1 cm based on ASE guidelines [16,20]. Measurements not available or not reported by either the AISAP software or the TTE report were noted and excluded from statistical analyses.

Weighted Cohen’s kappa (κ) coefficients were calculated for each of the ordinal categories to determine the level of agreement between AI POCUS and TTE. Weighted κ was chosen as it allows for partial agreement between adjacent severity categories, with penalties assigned to larger discrepancies. Bias-corrected and accelerated (BCa) bootstrap confidence intervals were calculated to account for non-normal data distribution and potential bias in κ estimates [22].

Agreement between LVEF and RVFAC measurements was further assessed using exploratory Bland-Altman analyses of representative numeric values mapped from the ordinal categories. Exact continuous LVEF and RVFAC values were not consistently available in formal TTE reports, which were frequently reported as ranges or categorical assessments. Therefore, LVEF categories in the study were mapped to 50%, 45%, 33%, and 25%, and RVFAC categories to 35%, 32.5%, 26%, and 22%, which represent the midpoint or boundary of each category. To assess the effect of assumptions for the open-ended categories, sensitivity analyses used alternative mappings of 55% and 20% for the highest (≥50%) and lowest (≤25%) LVEF categories, respectively, and 40% and 18% for the highest (>35%) and lowest (≤22%) RVFAC categories, respectively. Mean bias and limits of agreement (LoA) were calculated for both the primary and sensitivity mappings [23]. As an additional secondary analysis, percent agreement of AI POCUS with normal TTE findings was calculated and sensitivity and specificity at dichotomized thresholds for normal versus abnormal findings were calculated. Normal TTE findings were defined as LVEF ≥50%, RVFAC >35%, IVC ≤2.1 cm, or valvular regurgitation or stenosis of none to trace.

Statistical analysis scripts were generated using ChatGPT Enterprise (OpenAI GPT-5.5; accessed May 2026) in a configuration with model training disabled. All AI-generated code was reviewed, tested, and revised by the authors, who confirmed the accuracy and reproducibility of the results. Statistical analyses were performed using Python (version 3.13.5; Python Software Foundation). Data processing and statistical analyses were conducted using pandas (version 3.0.2), NumPy (version 2.4.4), SciPy (version 1.17.1), and scikit-learn (version 1.8.0).

Results

Patient Characteristics

We included a convenience sample of 70 participants in the analysis. Patient characteristics are shown in Table 1. There was no significant difference in patient characteristics between interpretable and non-interpretable scans when accounting for multiple comparisons (Supplementary Material S1 — Table S1). The median time between AI POCUS and TTE being performed was 1 day (IQR 1–2 days, total range 0–6 days).

Table 1. Baseline characteristics of study population.
CharacteristicValue
Patients, N70
Age, yr — mean ± SD68.5 ± 18.5
Age, yr — median (IQR)70.5 (61.3–82.8)
Male sex — N (%)42 (60)
Body-mass index, kg/m² — mean ± SD25.9 ± 5.0
Body-mass index, kg/m² — median (IQR)25.9 (22.6–28.6)
IQR, interquartile range; SD, standard deviation.

LVEF and RVFAC

The AI POCUS model provided an interpretation of LVEF in 45 patients (64%), while 25 patients had an indeterminate output from the model. Among these 45 individuals, there was 82% categorical agreement between AI POCUS and TTE, with an overall Cohen’s κ = 0.749 (95% CI 0.415–0.924; Figure 1 and Table 2). Exploratory Bland-Altman plots of LVEF with locked boundaries showed a mean bias of -0.67 (LoA -9.36–8.03) without marked change in mean bias but with slightly wider LoA under sensitivity analysis conditions, likely driven by a small number of outliers (Supplementary Material S1 — Figures S1A and S1B). For TTEs with normal LVEF, the AI POCUS model provided an interpretation in 60% of scans with 89% agreement (Figure 2; Supplementary Material S1 — Table S2). Using a dichotomized threshold of LVEF <50% versus ≥50%, AI POCUS demonstrated a sensitivity of 80.0% (95% CI 44.4–97.5%; Supplementary Material S1 — Table S3) and specificity of 88.6% (95% CI 73.3–96.8%; Supplementary Material S1 — Table S3) for detecting reduced LVEF.

Figure 1. Forest plot of weighted κ values for artificial intelligence (AI) point of care ultrasound (POCUS) versus transthoracic echocardiography (TTE) with 95% bias-corrected and accelerated (BCa) bootstrap confidence intervals. AR, aortic regurgitation; AS, aortic stenosis; IVC, inferior vena cava; LVEF, left ventricular ejection fraction; MR, mitral regurgitation; RVFAC, right ventricular fractional area change; TR, tricuspid regurgitation.

For RVFAC, the AI POCUS model provided an interpretation in 43 patients (61%), while 27 had an indeterminate output. Among those 43 individuals, there was 61% categorical agreement between AI POCUS and TTE, with an overall Cohen’s κ = -0.003 (95% CI -0.110–0.189; Figure 1 and Table 2). Exploratory Bland-Altman analysis with locked boundaries showed a mean bias of -1.67 and a LoA -9.79–6.45, with a larger mean bias of -3.02 and substantially wider LoA (-16.89–10.85) under sensitivity analysis conditions (Supplementary Material S1 — Figures S1C and S1D). For TTEs with normal RVFAC, the AI POCUS model provided an interpretation in 63% of scans with 65% agreement (Figure 2). Using a dichotomized threshold of RVFAC ≤35% versus >35%, AI POCUS demonstrated a sensitivity of 50.0% (95% CI 11.8–88.2%; Supplementary Material S1 — Table S3) and specificity of 64.9% (95% CI 47.5–79.8%; Supplementary Material S1 — Table S3) for detecting reduced RVFAC.

Table 2. Cohen’s κ agreement between artificial intelligence (AI) point of care ultrasound (POCUS) and transthoracic echocardiography (TTE).
ParameterPatients, NWeighted κ (95% BCa CI)Exact Agreement (%)
LVEF450.749 (0.415–0.924)82
MR280.438 (0.072–0.696)68
AR340.541 (0.155–0.841)85
AS250.086 (0.000–0.447)56
TR510.400 (0.144–0.634)71
IVC diameter540.382 (0.114–0.631)74
RVFAC43-0.003 (-0.110–0.189)61
Agreement was assessed using weighted Cohen’s κ, with 95% bias-corrected and accelerated (BCa) bootstrap confidence intervals. Weighted κ accounts for partial agreement between ordered categories. Exact agreement represents the percentage of observations classified identically by AI POCUS and TTE. AR, aortic regurgitation; AS, aortic stenosis; IVC, inferior vena cava; LVEF, left ventricular ejection fraction; MR, mitral regurgitation; RVFAC, right ventricular fractional area change; TR, tricuspid regurgitation.
Figure 2. Percent agreement of artificial intelligence (AI) point of care ultrasound (POCUS) with normal transthoracic echocardiography (TTE) findings by parameter. AR, aortic regurgitation; AS, aortic stenosis; IVC, inferior vena cava; LVEF, left ventricular ejection fraction; MR, mitral regurgitation; RVFAC, right ventricular fractional area change; TR, tricuspid regurgitation.

IVC Diameter

For subcostal IVC evaluation, the AI POCUS model provided an interpretation in 54 patients (77%), while 16 had an indeterminate output. Among those 54 individuals, there was 74% categorical agreement between AI POCUS and TTE, with an overall Cohen’s κ = 0.382 (95% CI 0.114–0.631; Figure 1 and Table 2). For TTEs with normal IVC diameter, the AI POCUS model provided an interpretation in 77% of scans with 78% agreement (Figure 2). Using a dichotomized threshold of IVC >2.1 cm versus ≤2.1 cm, AI POCUS demonstrated a sensitivity of 64.3% (95% CI 35.1–87.2%; Supplementary Material S1 — Table S3) and specificity of 77.5% (95% CI 61.5–89.2%; Supplementary Material — Table S3) for detecting an IVC >2.1 cm.

Valvular Evaluation

For AR, the AI POCUS model provided interpretation in 34 patients (49%), while 36 had indeterminate output. Among those 34 individuals, there was 85% categorical agreement between AI POCUS and TTE, with an overall Cohen’s κ = 0.541 (95% CI 0.155–0.841; Figure 1 and Table 2). For TTEs with AR of none to trace, the AI POCUS model provided an interpretation in 47% of scans with 100% agreement (Figure 2). AI POCUS demonstrated a sensitivity of 44.4% (95% CI 13.7–78.8%; Supplementary Material S1 — Table S3) and specificity of 100% (95% CI 86.3–100%; Supplementary Material S1 — Table S3) for detecting greater than trace AR.

For MR, the AI POCUS model provided interpretation in 28 patients (40%), while 42 had indeterminate output. Among those 28 individuals, there was 68% categorical agreement between AI POCUS and TTE, with an overall Cohen’s κ = 0.438 (95% CI 0.072–0.696; Figure 1 and Table 2). For TTEs with MR of none to trace, the AI POCUS model provided an interpretation in 36% of scans with 71% agreement (Figure 2). AI POCUS demonstrated a sensitivity of 72.7% (95% CI 39.0–94.0%; Supplementary Material S1 — Table S3) and specificity of 70.6% (95% CI 44.0–89.7%; Supplementary Material S1 — Table S3) for detecting greater than trace MR.

For TR, the AI POCUS model provided interpretation in 51 patients (73%), while 19 had indeterminate output. Among those 51 individuals, there was 71% categorical agreement between AI POCUS and TTE, with an overall Cohen’s κ = 0.400 (95% CI 0.144–0.634; Figure 1 and Table 2). For TTEs with TR of none to trace, the AI POCUS model provided an interpretation in 68% of scans with 85% agreement (Figure 2). AI POCUS demonstrated a sensitivity of 54.2% (95% CI 32.8–74.4%; Supplementary Material S1 — Table S3) and specificity of 85.2% (95% CI 66.3–95.8%; Supplementary Material S1 — Table S3) for detecting greater than trace TR.

For AS, the AI POCUS model provided interpretation in 25 patients (36%), while 45 had indeterminate output. Among those 25 individuals, there was 56% categorical agreement between AI POCUS and TTE, with an overall Cohen’s κ = 0.086 (95% CI 0.000–0.447; Figure 1 and Table 2). For TTEs with AS of none to trace, the AI POCUS model provided an interpretation in 41% of scans with 54% agreement (Figure 2). AI POCUS demonstrated a sensitivity of 100% (95% CI 2.5–100%; Supplementary Material S1 — Table S3) and specificity of 54.2% (95% CI 32.8–74.4%; Supplementary Material S1 — Table S3) for detecting greater than trace AS.

Discussion

In this single-arm, prospective diagnostic comparison study, we compared a commercially available AI POCUS model with standard TTE performed within a median of one day among hospitalized patients. We observed strong agreement for LVEF between AI POCUS and TTE, while there was lower agreement for IVC diameter, valvular assessments, and RVFAC. When the TTE reported normal parameters, the AI POCUS model had high levels of agreement for LVEF, AR, TR, and IVC diameter. AI POCUS demonstrated moderate-to-high sensitivity for detecting reduced LVEF and greater than trace MR. At the dichotomized thresholds used in this study, specificity generally exceeded sensitivity, which suggests more consistent classification of normal findings than mild or greater abnormalities. However, sensitivity estimates were imprecise because of the small number of abnormal cases for several parameters and should not be interpreted as evidence that AI POCUS is uniformly unable to detect clinically significant disease. These findings highlight both the strengths and current limitations of AI POCUS compared with standard TTE and support its potential role as an adjunctive screening tool rather than a replacement for definitive echocardiographic diagnosis and management. Our study builds on existing literature by evaluating AI POCUS prospectively and in a clinical context, providing a pragmatic assessment of agreement between imaging modalities.

The differences in agreement between LVEF and other parameters likely reflect both image acquisition and physiologic factors. LVEF can be rapidly and reliably estimated with POCUS using PLAX and PSAX views, which are easier and faster to obtain compared to the A4C view required to calculate RVFAC [24–26]. In contrast, the RV has a complex, crescentic shape and sits behind the sternum, making it more challenging to assess in the A4C acoustic window [27]. RVFAC is particularly affected by image foreshortening [20], which is common among POCUS users and contributes to reduced reproducibility even in standard echocardiography [28–30]. For IVC, the measured diameter may vary based on volume status, respiratory status, and patient positioning, which may lead to moderate interrater reliability [31]. This observed variability in IVC may have impacted our study with the additional factor of the time difference between when POCUS and the TTE scans were obtained. Lastly, valvular quantifications are some of the most challenging tasks in echocardiography, requiring multiple technical inputs with particular difficulty in differentiating mild from moderate severity. Given that the AI tool uses colour Doppler to make measurements without other standard assessments, this could have led to lower agreement with TTE [32].

Our study builds on prior work evaluating AI POCUS applications, many of which focus on model development and classification performance based on retrospective datasets [17,18]. Fisher et al. applied the AI deep learning model from this study retrospectively to over 120,000 TTE exams, demonstrating that a single A4C view could accurately identify valvular disease and RV dysfunction while reliably estimating LVEF [18]. In another retrospective study, Cohen et al. demonstrated high diagnostic accuracy of the AI model in identifying clinically significant MR (AUC 0.98) and TR (AUC 0.96) in formal TTE images [17]. Notably, in the Cohen et al. study, predictions were generated in approximately 38% of studies, suggesting the sensitivity of the model to input image quality [17].

Interpretability in our study likewise varied substantially by parameter, ranging from 36% for AS to 77% for IVC diameter, and was in several instances (LVEF, RVFAC, AS, AR, and MR) lower than the rates recently reported by Fisher et al. in a separate prospective study using the same AI platform and ultrasound probe [33]. Several factors may help explain this difference. Novice operators in the Fisher et al. study completed a structured 6-hour training program with cardiologist supervision of each operator’s first 20 examinations and ongoing access to real-time expert support, an infrastructure not available to the more experienced POCUS operators in our study. The AISAP software also provides a real-time image-quality indicator that requires a consistent internet connection; this feature was used in the Fisher et al. study but was not consistently available on the devices used in our study, which may have favoured acquisition of higher-quality images in their cohort. Fisher et al. was also conducted in a different country (Israel), and differences in patient population or disease prevalence could have further contributed to the observed differences in interpretability. Because the studies were not designed to directly compare acquisition workflows, we cannot quantify the individual contribution of these factors.

Another prospective investigation by Fisher et al. using the AI POCUS software utilized in our study demonstrated strong performance of AI POCUS for detecting reduced LVEF (AUC 0.97) and reduced RV function (AUC 0.95), with more modest performance for detecting at least moderate MR (AUC 0.72) and TR (AUC 0.86) based on a single A4C view, compared to standard TTE [18]. These findings underscore an important distinction between diagnostic discrimination and clinical interchangeability. Although the AI system may reliably categorize normal versus abnormal findings, this does not necessarily imply agreement with comprehensive echocardiography across a full range of measurements. In our study, we build on this work by directly evaluating agreement between AI POCUS and formal TTE across multiple measurements and severity categories, demonstrating strong concordance for LVEF, with lesser agreement for RVFAC, IVC diameter, and valvular assessments. Our study adds to the literature supporting AI POCUS as a powerful tool for identifying and categorizing LVEF, while highlighting its current limitations in replacing comprehensive TTE.

Diagnostic thresholds also differed across studies, which complicates direct comparison of sensitivity and specificity. Our prespecified analysis was consistent with ASE guidelines and classified any abnormality as positive (LVEF <50%, RVFAC ≤35%, IVC diameter >2.1 cm, and valvular disease greater than none-to-trace). Fisher et al. defined clinically significant disease as LVEF <40% and moderate-or-greater valvular or RV dysfunction [33]. We therefore repeated our LVEF analysis at the closest comparable threshold supported by our categorical data (LVEF ≤40%); sensitivity was 80.0% (95% CI, 28.4% to 99.5%) and specificity was 95.0% (95% CI, 83.1% to 99.4%), compared with 80.0% and 88.6%, respectively, at the <50% threshold (Supplementary Material S1 — Table S4). Specificity at this threshold approximated that reported by Fisher et al., although sensitivity remained lower and was based on only five TTE-positive cases. An equivalent ≥moderate valvular reanalysis was not feasible, as our mild-to-moderate and moderate-to-severe categories cross the ≥moderate cutoff and cannot be retrospectively separated without introducing classification assumptions. These comparisons are further limited by differences in study design: the noninterventional AI POCUS validation study drew on 5,203 formal echocardiographic examinations rather than bedside POCUS, and the Fisher et al. and Cohen et al. estimates were calculated only among patients selected for confirmatory echocardiography rather than a systematically sampled cohort, introducing potential verification bias of indeterminate direction and magnitude. Prior studies also reported point estimates without confidence intervals. Given our smaller sample, several of our confidence intervals are wide, and cross-study differences in point estimates for sensitivity and specificity should be interpreted in this context rather than as direct measures of relative model performance [17,18,33].

There are several implications of this study in clinical practice. Given the strong agreement in LVEF in this study between AI POCUS and TTE, it could be used as a screening tool for LVEF in patients presenting with symptoms of heart failure, which is supported by prior investigations that found that AI systems perform well in identifying reduced LVEF [18,25,33,34]. Similarly, AI POCUS could be used to screen for clinically significant valvular disease [11]. Employing a point-of-care tool such as AI POCUS that can reliably detect normal versus low LVEF or screen for the presence of more severe valvular dysfunction on a population level could help determine which patients should undergo formal TTE or hospitalization for further testing, particularly in low-resource settings. It is worth emphasizing that AI POCUS has not demonstrated the capability to replace TTE as a standard assessment.

There are several limitations to this study. This was a single-institution study derived from a convenience sample of inpatient participants. Each participant underwent a single scan by a member of the research team, and interrater reliability was not calculated among POCUS scanners. Although the median time from POCUS scan to TTE was one day, variations in echocardiographic findings, especially in dynamic inpatient settings where patients are acutely ill, may have contributed to differences between the two studies. This is especially true for IVC diameter and valvular function, which may vary with changes in volume status [35]. The overall sample size was modest, with only fifteen AI POCUS scans resulting in complete imaging reports. The remaining scans had missing measurements due to inability to obtain the required view or the view classification feature deeming the inputs unsuitable for the diagnostic model. This may have led to limited precision for less prevalent findings, particularly AS, MR, and AR, but is also reflective of POCUS in clinical practice, where adequate views may be difficult to obtain in some patients. Successful AI interpretation depends on acquisition of adequate quality images and the required echocardiographic views, and inadequate image quality may prevent the platform from generating a diagnostic interpretation. In particular, the apparent 100% sensitivity for detecting AS (Supplementary Material S1 — Table S3) was derived from a small number of interpretable and TTE-positive cases, yielding a correspondingly wide confidence interval, and should not be interpreted as evidence of reliable AS detection by AI POCUS. Rather, the small number of cases limits conclusions regarding AS performance and underscores the need for caution when integrating AI POCUS findings into clinical decision-making. Lastly, all scans were obtained using a single device (Philips Lumify S4-1 Phased Array Probe) and analyzed by a single commercial AI product, limiting generalizability to other devices and AI systems.

Conclusions

In this single-arm, prospective diagnostic comparison study, AI POCUS demonstrated close agreement with standard TTE for the estimation of LVEF. Agreement was lower for RVFAC, valvular assessments, and IVC diameter. These findings support the potential role of AI POCUS for assessing left ventricular function, particularly where access to comprehensive echocardiography is limited. Further integration and validation studies are needed to assess the role of AI POCUS in clinical practice and its impact on management decisions.

Ethics Statement

This study received human subjects approval by the Stanford University Institutional Review Board (IRB Protocol #77758). The requirement for informed consent was waived by the IRB.

Disclosure Statement

AK serves as a consultant for GE Healthcare and Exo Inc. EB is a co-investigator for a study sponsored by GE Healthcare.

Funding

No associated funding was received for this study

Author Contributions

EB: conceptualization, data curation, methodology, formal analysis, writing – original draft, writing – review & editing. JK: data curation, project administration, writing – original draft, writing – review & editing. ML: data curation, writing – original draft, writing – review & editing. CPC: data curation, writing – original draft, writing – review & editing. VN: data curation, writing – original draft, writing – review & editing. AK: conceptualization, data curation, methodology, supervision, writing – original draft, writing – review & editing.

References

1. Savarese G, Becher PM, Lund LH, Seferovic P, Rosano GMC, Coats AJS. Global burden of heart failure: a comprehensive and updated review of epidemiology. Cardiovasc Res 2023; 118(17):3272-3287.

2. Global Burden of Cardiovascular Diseases and Risks 2023 Collaborators. Global, regional, and national burden of cardiovascular diseases and risk factors in 204 countries and territories, 1990-2023. J Am Coll Cardiol 2025; 86(22):2167-2243.

3. Heidenreich PA, Bozkurt B, Aguilar D, Allen LA, Byun JJ, Colvin MM, Deswal A, Drazner MH, Dunlay SM, Evers LR, Fang JC, Fedson SE, Fonarow GC, Hayek SS, Hernandez AF, Khazanie P, Kittleson MM, Lee CS, Link MS, Milano CA, Nnacheta LC, Sandhu AT, Stevenson LW, Vardeny O, Vest AR, Yancy CW. 2022 AHA/ACC/HFSA guideline for the management of heart failure: Executive summary: A report of the American College of Cardiology/American Heart Association Joint Committee on Clinical Practice Guidelines. J Am Coll Cardiol 2022; 79(17):1757-1780.

4. Yadav H, Shah D, Sayed S, Horton S, Schroeder LF. Availability of essential diagnostics in ten low-income and middle-income countries: results from national health facility surveys. Lancet Glob Health 2021; 9(11):e1553-e1560.

5. Díaz-Gómez JL, Mayo PH, Koenig SJ. Point-of-care ultrasonography. N Engl J Med 2021; 385(17):1593-1602.

6. Qaseem A, Etxeandia-Ikobaltzeta I, Mustafa RA, Kansagara D, Fitterman N, Wilt TJ, Clinical Guidelines Committee of the American College of Physicians, Batur P, Cooney TG, Crandall CJ, Hicks LA, Lin JS, Maroto M, Tice J, Tufte JE, Vijan S, Williams JW Jr. Appropriate use of point-of-care ultrasonography in patients with acute dyspnea in emergency department or inpatient settings: A clinical guideline from the American College of Physicians. Ann Intern Med 2021; 174(7):985-993.

7. Maganti K, Chen C, Jamthikar AD, Parikh P, Yanamala N, Sengupta PP. Cardiopulmonary point-of-care ultrasonography for hospitalist management of undifferentiated dyspnea. JAMA Netw Open 2025; 8(9):e2530677.

8. Baum E, Tandel MD, Ren C, Weng Y, Pascucci M, Kugler J, Cardoza K, Kumar A. Acquisition of cardiac point-of-care ultrasound images with deep learning: A randomized trial for educational outcomes with novices. CHEST Pulm 2023; 1(3):100023.

9. Trost B, Rodrigues L, Ong C, Dezellus A, Goldberg YH, Bouchat M, Roger E, Moal O, Singh V, Moal B, Lafitte S. Artificial intelligence empowers novice users to acquire diagnostic-quality echocardiography. JACC Adv 2025; 4(8):102005.

10. Kumar A, Baum E, Parmer-Chow C, Kugler J. Limited echocardiogram acquisition by novice clinicians aided with deep learning: A randomized controlled trial. Biol Methods Protoc 2025; 10(1):bpaf083.

11. Holste G, Oikonomou EK, Tokodi M, Kovács A, Wang Z, Khera R. Complete AI-enabled echocardiography interpretation with multitask deep learning. JAMA 2025; 334(4):306-318.

12. Kagiyama N, Abe Y, Kusunose K, Kato N, Kaneko T, Murata A, Ota M, Shibayama K, Izumo M, Watanabe H. Multicenter validation study for automated left ventricular ejection fraction assessment using a handheld ultrasound with artificial intelligence. Sci Rep 2024; 14(1):15359.

13. Motazedian P, Marbach JA, Prosperi-Porta G, Parlow S, Di Santo P, Abdel-Razek O, Jung R, Bradford WB, Tsang M, Hyon M, Pacifici S, Mohanty S, Ramirez FD, Huggins GS, Simard T, Hon S, Hibbert B. Diagnostic accuracy of point-of-care ultrasound with artificial intelligence-assisted assessment of left ventricular ejection fraction. NPJ Digit Med 2023; 6(1):201.

14. Vega R, Dehghan M, Nagdev A, Buchanan B, Kapur J, Jaremko JL, Zonoobi D. Overcoming barriers in the use of artificial intelligence in point of care ultrasound. NPJ Digit Med 2025; 8(1):213.

15. Lin JC, Jain B, Iyer JM, Rola I, Srinivasan AR, Kang C, Patel H, Parikh RB. Benefit-risk reporting for FDA-cleared artificial intelligence-enabled medical devices. JAMA Health Forum 2025; 6(9):e253351.

16. Mitchell C, Rahko PS, Blauwet LA, Canaday B, Finstuen JA, Foster MC, Horton K, Ogunyankin KO, Palma RA, Velazquez EJ. Guidelines for performing a comprehensive transthoracic echocardiographic examination in adults: Recommendations from the American Society of Echocardiography. J Am Soc Echocardiogr 2019; 32(1):1-64.

17. Cohen I, Malins JG, Cohen-Shelly M, Asaf Y, Fiman M, Faierstein K, Fisher L, Sudri K, Raanani E, Schwammenthal E, Klempfner R, Maor E. Deep learning for atrioventricular regurgitation diagnosis: an external validation study. Eur Heart J Digit Health 2025; 6(5):949-958.

18. Fisher L, Fiman M, Segal E, Lidar S, Rubin N, Am-Shalom A, Cohen I, Faierstein K, Tsur AM, Schwammenthal E, Klempfner R, Zimlichman E, Raanani E, Maor E. Artificial intelligence assessment of valvular disease and ventricular function by a single echocardiography view. Front Digit Health 2025; 7:1684933.

19. Doherty JU, Daugherty SL, Kort S, London MJ, Mehran R, Merli GJ, Schoenhagen P, Soman P, Starling RC, Johnson DM, Dehmer GJ, Bhave NM, Biederman RW, Bittencourt MS, Burroughs MS, Doukky R, Hays AG, Indik JH, Kim KM, Lotfi AS, Macchiavelli AJ, Neuburger P, Patel H, Pellikka PA, Reece TB, Rong LQ. ACC/AHA/ASE/ASNC/HFSA/HRS/SCAI/SCCT/SCMR/STS 2024 appropriate use criteria for multimodality imaging in cardiovascular evaluation of patients undergoing nonemergent, noncardiac surgery. J Am Coll Cardiol 2024; 84(15):1455-1491.

20. Mukherjee M, Rudski LG, Addetia K, Afilalo J, D’Alto M, Freed BH, Friend LB, Gargani L, Grapsa J, Hassoun PM, Hua L, Kim J, Mercurio V, Saggar R, Vonk-Noordegraaf A. Guidelines for the echocardiographic assessment of the right heart in adults and special considerations in pulmonary hypertension: Recommendations from the American Society of Echocardiography. J Am Soc Echocardiogr 2025; 38(3):141-186.

21. Otto CM, Nishimura RA, Bonow RO, Carabello BA, Erwin JP 3rd, Gentile F, Jneid H, Krieger EV, Mack M, McLeod C, O’Gara PT, Rigolin VH, Sundt TM 3rd, Thompson A, Toly C. 2020 ACC/AHA guideline for the management of patients with valvular heart disease: A report of the American College of Cardiology/American Heart Association Joint Committee on Clinical Practice Guidelines. Circulation 2021; 143(5):e72-e227.

22. Efron B. Better bootstrap confidence intervals. J Am Stat Assoc 1987; 82(397):171-185.

23. Bland JM, Altman DG. Statistical methods for assessing agreement between two methods of clinical measurement. Lancet 1986; 1(8476):307-310.

24. Lebeau R, Robert-Halabi M, Pichette M, Vinet A, Sauvé C, Dilorenzo M, Le V, Piette E, Brunet M, Bédard W, Serri K, Poulin F. Left ventricular ejection fraction using a simplified wall motion score based on mid-parasternal short axis and apical four-chamber views for non-cardiologists. BMC Cardiovasc Disord 2023; 23(1):115.

25. Gallant C, Bernard L, Kwok C, Wichuk S, Noga M, Punithakumar K, Hareendranathan A, Becher H, Buchanan B, Jaremko JL. AI-augmented point of care ultrasound in intensive care unit patients: Can novices perform a “basic echo” to estimate left ventricular ejection fraction in this acute-care setting? J Clin Med 2025; 14(9):2899.

26. Bahl A, Johnson S, Altwail M, Brackney A, Xiao J, Price J, Shotkin P, Chen NW. Left ventricular ejection fraction assessment by emergency physician-performed bedside echocardiography: A prospective comparative evaluation of multiple modalities. J Emerg Med 2021; 61(6):711-719.

27. Hahn RT, Lerakis S, Delgado V, Addetia K, Burkhoff D, Muraru D, Pinney S, Friedberg MK. Multimodality imaging of right heart function: JACC scientific statement. J Am Coll Cardiol 2023; 81(19):1954-1973.

28. Spencer KT, Flachskampf FA. Focused cardiac ultrasonography. JACC Cardiovasc Imaging 2019; 12(7 Pt 1):1243-1253.

29. Mor-Avi V, Khandheria B, Klempfner R, Cotella JI, Moreno M, Ignatowski D, Guile B, Hayes HJ, Hipke K, Kaminski A, Spiegelstein D, Avisar N, Kezurer I, Mazursky A, Handel R, Peleg Y, Avraham S, Ludomirsky A, Lang RM. Real-time artificial intelligence-based guidance of echocardiographic imaging by novices: Image quality and suitability for diagnostic interpretation and quantitative analysis. Circ Cardiovasc Imaging 2023; 16(11):e015569.

30. Marwick TH. Ejection fraction pros and cons: JACC state-of-the-art review. J Am Coll Cardiol 2018; 72(19):2360-2379.

31. Policastro P, Ermini L, Civera S, Albani S, Musumeci G, Roatta S, Mesin L. Effect of respirophasic displacement of the inferior vena cava on size measurement in 2-D ultrasound imaging. Ultrasound Med Biol 2024; 50(12):1785-1792.

32. Zoghbi WA, Adams D, Bonow RO, Enriquez-Sarano M, Foster E, Grayburn PA, Hahn RT, Han Y, Hung J, Lang RM, Little SH, Shah DJ, Shernan S, Thavendiranathan P, Thomas JD, Weissman NJ. Recommendations for noninvasive evaluation of native valvular regurgitation: A report from the American Society of Echocardiography developed in collaboration with the Society for Cardiovascular Magnetic Resonance. J Am Soc Echocardiogr 2017; 30(4):303-371.

33. Fisher L, Fiman M, Yarkoni Y, Segal E, Fishman B, Faierstein K, Rubin N, Am-Shalom A, Amital H, Zhao Q, Kort S, Mehrotra P, Schwammenthal E, Klempfner R, Zimlichman E, Raanani E, Maor E. Artificial intelligence-enhanced cardiac point-of-care ultrasound: A prospective single-arm study. Mayo Clin Proc Digit Health 2026; 4(2):100355.

34. Campbell RT, Petrie MC, Docherty KF, Brooksbank KJM, McKinley G, Haig C, McConnachie A, Lam CSP, Adamson C, Butler E, Curtain JP, Hartshorne-Evans N, Graham FJ, Hainey H, Jarvie J, Lee MMY, Macklin L, Mangion K, McCoubrey A, McDowell K, McIntyre A, Nordin S, Osmanska J, Pellicori P, Simpson J, Sonecki P, Taylor K, Taylor-Sweet D, Turnbull P, Welsh P, McMurray JJV, Murphy CL, Lowe DJ. Artificial intelligence fully automated analysis of handheld echocardiography in real-world patients with suspected heart failure. Eur J Heart Fail 2025; 27(12):3401-3410.

35. Kawata T, Daimon M, Nakanishi K, Kimura K, Sawada N, Nakao T, Hirokawa M, Ishiwata J, Takeda N, Yatomi Y, Komuro I. Factors influencing inferior vena cava diameter and its respiratory variation: Simultaneous comparison with hemodynamic data. J Cardiol 2022; 79(5):642-647.

Posted in .

Leave a Reply

Your email address will not be published. Required fields are marked *